Tiny Reservoir Computing for Extreme Learning of Motor Control

The electromechanical system of an electric machine controller represents a complex non-linear system that needs to adapt itself continuously over time, since factors as component aging and thermal derating tend to modify its behavior and performance. Algorithms based on Neural Networks (NN), and in particular Recurrent NN (RNN), are expected to improve the motor drive, due to their ability to approximate complex non-linear systems, and to cope with the time-varying nature of signals. This paper proposes a novel control technique which extends the Field Oriented Control (FOC) algorithm by means of Extreme Learning Machine (ELM) and Reservoir Computing (RC). In particular, we introduce a specifically developed NN, named Semi Binary Deep Echo State Networks (SB-DESN), to achieve good control accuracy with low complexity. This allows its deployment into cheap micro-controllers characterized by severe memory, computational, and power consumption constraints. The peculiar property of SB-DESN i.e., a fixed and predictable training complexity, makes it well suited for adoption into Reinforcement Learning on micro-controllers to ensure the continuous adaptability of the control over the time. Beyond the use of SB-DESN for speed, torque, and magnetic flux control, we propose a novel complexity optimization based on sparsity properties of inter-reservoir matrices, which reduces the memory footprint up to a factor 3.7 and the computational complexity up to a factor 2. This corresponds to a SB-DESN with only 2.5kB of memory footprint and an inference time of 20µs on STM32H7, which allows its deployment into mass market FOC solution.